Subject-Specific Residual Variability

Why Let Residual Variability Differ by Subject?

A population model often allows subjects to differ in pharmacokinetic parameters, such as clearance or volume. It may also allow concentrations from some subjects to be more dispersed around their own predicted concentration-time profiles than concentrations from other subjects. This second type of heterogeneity is subject-specific residual variability. It concerns the scale of the remaining observation error, not a subject’s typical pharmacokinetic trajectory.

The distinction matters when repeated concentrations from a subject repeatedly lie close to that subject’s predictions, or repeatedly show larger deviations, beyond what one common residual variance would explain. A single large deviation is not by itself evidence of a high-variability subject: it could simply be an ordinary residual draw, an outlier, or a poor prediction of the mean.

A Random Effect on the Residual Scale

Let \(Y_{ij}\) be the concentration for subject \(i\) at observation \(j\), and let \(f_{ij}(\theta,\eta_i)\) be the predicted concentration from the structural model. Here, \(\theta\) denotes population parameters and \(\eta_i\) denotes the subject’s pharmacokinetic random effects. One simple model places a random effect \(b_i\) on the logarithm of the residual standard deviation:

In this formulation, \(\sigma\) is the baseline residual standard deviation and \(\tau^2\) is the variance of the subject-level log-scale effect. The conditional standard deviation for subject \(i\) is \(\sigma_i=\sigma\exp(b_i)\); the conditional variance is \(\sigma_i^2=\sigma^2\exp(2b_i)\). Thus, \(b_i=0\) corresponds to the baseline scale, and a one-unit increase in \(b_i\) multiplies the standard deviation by \(\exp(1)\). Since \(b_i\) is normally distributed with mean zero, \(\sigma\) is the median subject-specific standard deviation under this model, not its population mean.

The model assumes the residuals have conditional mean zero and, given the random effects, are normally distributed and independent across observations. It also assumes the residual-scale effect is shared across observations for a subject, and that \(b_i\) is independent of the pharmacokinetic effects \(\eta_i\), unless a different dependence structure is explicitly specified. These are modeling assumptions, not automatic properties of concentration data. A known concentration-dependent residual variance function can be used in place of the constant baseline \(\sigma^2\); the subject-level multiplier then scales that variance function.

The residual model may instead be specified on a transformed scale, such as log concentration. In that case the observation, prediction, and residual variance are defined on that scale. A random effect on residual scale should not be confused with a random effect that changes the mean on the original concentration scale.

What the Random Effect Implies Across Observations

Although residuals from the same subject are conditionally independent in the stated model, their squared values are associated after accounting for the unobserved shared \(b_i\). If one residual from a subject is large in magnitude, that provides information about the subject’s residual scale and therefore about the expected magnitude of the subject’s other residuals. The signed residuals need not have nonzero covariance: conditional means remain zero.

The following derivation makes the role of repeated observations explicit. Write \(R_{ij}=Y_{ij}-f_{ij}(\theta,\eta_i)\). Given \(b_i\), distinct residuals are independent, and \(E(R_{ij}^2\mid b_i)=\sigma^2\exp(2b_i)\). Therefore:

For \(b_i\sim N(0,\tau^2)\), the normal moment-generating function gives \(E[\exp(tb_i)]=\exp(t^2\tau^2/2)\). Applying this result at \(t=2\) and \(t=4\) yields the population mean and variance of the subject-specific residual variance:

Consequently, for distinct observations from the same subject, the covariance of their squared residuals is \(\sigma^4\{\exp(8\tau^2)-\exp(4\tau^2)\}\). It is zero when \(\tau^2=0\) and positive when \(\tau^2>0\). This is a dependence in residual magnitudes induced by a shared scale effect, not a serial correlation in signed residuals.

Why Replication Within Subjects Matters

A subject-specific residual variance describes a subject’s dispersion, which is a feature of a collection of observations rather than of one signed residual. Repeated observations allow the model to compare how widely a subject’s residuals vary with how widely residuals vary for other subjects. For this comparison to be informative, residuals must be calculated relative to adequate predictions: systematic mean-model misspecification can look like excess residual dispersion.

With only one observation per subject, there is no within-subject sample variance. Across many singleton subjects, a model with a specified residual distribution may still contain information about \(\tau^2\) through the shape of the marginal residual distribution. But that information depends heavily on distributional assumptions, and an unusually large residual cannot be separated cleanly into a high subject-specific scale, a chance residual draw, or a mean-model problem. Repeated measurements provide direct within-subject evidence and are generally important for reliable estimation and interpretation.

Distinguishing Residual-Scale Effects From Other Effects

A random effect on a pharmacokinetic parameter changes the predicted profile. For example, a subject-specific clearance effect can alter the predicted concentrations over time, and its influence is structured by the pharmacokinetic model. A residual-scale effect instead leaves the conditional predicted profile unchanged and changes the expected dispersion around it. If a subject’s residuals consistently shift in one direction or follow a time pattern, a scale effect alone is not an adequate explanation.

An occasion effect is also distinct. It is attached to an occasion or dosing period and can change a parameter or residual scale for observations in that occasion. A subject-level residual effect is shared across that subject’s observations, including across occasions if that is how the model is defined. The level at which an effect is shared must match the scientific question and sampling design.

The main alternatives are to retain one common residual scale, to use a fixed variance function that changes with predicted concentration, or to specify occasion-specific residual variances. A common scale is simpler but cannot represent persistent between-subject differences in dispersion. A fixed variance function describes systematic changes in spread with a measured predictor or predicted concentration, not unexplained subject-to-subject scale differences. Occasion-specific variances allow group-level differences between occasions, but do not by themselves represent individual subjects having persistently different residual scales.

Diagnostics, Failure Modes, and Limitations

Inspect residual patterns at both the subject and population levels. If subjects with large residual magnitudes tend to have large magnitudes repeatedly, the shared-scale idea may be relevant. Plotting residuals against predictions and time can reveal a variance trend, a systematic mean pattern, or an occasion effect that a subject-specific scale cannot explain. Assess whether a few extreme observations dominate the apparent heterogeneity; a heavy-tailed residual distribution or data-quality issue can be mistaken for subject-specific variability.

The model also has limits. It assumes a chosen distribution for the subject-level effect and a chosen conditional residual distribution. A lognormal scale distribution permits very large scales, so a small number of subjects can have substantial influence. Sparse sampling weakens the information about each subject’s dispersion, while a misspecified mean model can inflate the estimated residual-scale variability. If the subject-scale effect is correlated with pharmacokinetic random effects, assuming independence may also distort estimates; that dependence must be considered explicitly rather than inferred from residual spread alone.

A common mistake is to call a subject with one large residual a “high residual-variance subject.” One residual supplies evidence about a particular error draw, not a direct estimate of a subject’s variance. The correction is to distinguish the residual realization from the latent subject-level scale, use repeated observations to assess persistent dispersion, and check that mean predictions and other variance patterns are adequate before attributing the spread to this random effect.